Graph Similarity Computation Using Convolution Neural network

Abstract

Graph similarity computation is a foundational problem with broad applications including molecule comparison, social network analysis, bioinformatics, and code analysis. Traditional methods such as GED and MCS are NP-hard and impractical for large graphs. The emergence of graph neural networks (GNNs), particularly convolution-based architectures has revolutionized graph similarity learning by leveraging deep learning techniques for efficient and scalable prediction. This paper provides a comprehensive overview of convolutional neural network (CNN)–based architectures for graph similarity computation, emphasizing their scalability, interpretability, and adaptability. The discussion extends to graph autoencoders (GAE), generative adversarial networks (GAN), and hybrid frameworks that learn structural embeddings capable of generalizing across domains. We analyze key models including SimGNN, GraphSim, GSimCNN, and SEGMN, highlighting their mechanisms for multi-scale feature extraction, convolutional matching, and structure-enhanced reasoning. This review comprehensively examines convolutional architectures for graph similarity computation, including mathematical foundations, CNN-based set matching, node ordering schemes, applications, challenges and future research directions.